Axiom BOINC Session Results Log
Session timestamp: 2026-03-03 21:21:40 (America/Denver)
Scope: Combined Part 1 (validation/credit) + Part 2 (research/deployment)

PART 1 - VALIDATION / CREDIT SUMMARY
- Reviewed uncredited failed/error cohort and diagnostics; sampled 240 uncredited completed rows for upload payload presence.
- Payload availability in sample was 0/240 (no experiment_result or error JSON artifacts found).
- Awarded credit to 10,000 completed exp_* rows at 1.0 credit each (session cap reached exactly: 10,000 total credit).
- Credited batch outcome mix: success=0, failed/error=10,000 (legacy zero-time failure dominated).
- Attribution split: 2,442 credit attributable to resolved hosts/users; 7,558 credit on rows with unresolved host metadata (result-level credit granted).
- Top attributable user additions: ChelseaOilman +407, makracz +217, Josemi +179, mmonnin +146, Steve Dodd +124, Orange Kid +104.
- Post-session backlog: uncredited success=0, uncredited failed/error=12,106.
- Website counters after update: credited_count=64,451; total_results_count=55,009.

STUCK / BROKEN TASK CLEANUP
- Dead-host >12h cleanup aborts: 0
- Hard-ceiling >48h cleanup aborts: 0
- Broken-prefix diagnostic: exp_gpu_matmul_stress_gpu* had 9 recent zero-time failures across 4 hosts, but active unsent/in-progress rows were 0, so no abort action was needed.

PART 2 - DEPLOYMENT / RESEARCH SUMMARY
- Part 2 started at 2026-03-03 14:03:44 (America/Denver) per AutoReview run log.
- Retirement pass executed against over-completed prefixes; ABORT_TOTAL=0 (no unsent retirement candidates to abort).

CPU DEPLOYMENT
- CPU scripts used for fill pass: wd_batchnoise_interaction.py, wd_labelsmooth_interaction.py.
- CPU deployment counters from run log: CPU_HOSTS_SEEN=81, CPU_SKIPPED_LOW_RAM=2, CPU_WU_CREATED=2937.
- CPU host targeting used host-specific workunit naming (_h<hostid>) and fill-to-queue logic (up to 3x CPU slots).
- CPU host coverage observed for new wd_batchnoise_interaction prefix: 24 hosts (1,144 rows observed).
- CPU host list (wd_batchnoise_interaction observed):
  287(DESKTOP-N5RAJSE), 345(Andre-WEBK), 340(Foxtrot-3), 333(Golf-2), 339(Foxtrot-2), 85(DadOld-PC), 159(achernar), 335(Hotel-3), 327(Echo-3), 332(Delta-2), 330(Delta-1), 123(Dads-PC), 1(Pyhelix), 324(Charlie-1), 321(Rosie), 29(DESKTOP-P57624Q), 342(PC-1), 87(Dad-Workstation), 222(DESKTOP-11MAEMP), 341(DESKTOP-ELBSBOI), 347(sirius), 352(procyon), 7(iand-r7-5800h3), 6(iand-r7-5800h).

GPU CHECKPOINT (MANDATORY)
- GPU deployment pass was started with scripts: wd_curvature_trigger_gpu.py and wd_timing_scale_gpu.py.
- The captured run_2026-03-03_1352.log ends with '^C' during GPU deployment script execution, so final in-log GPU creation counters were not recorded in that log file.
- Post-run server snapshot for these GPU prefixes shows:
  - GPU hosts observed: 18
  - GPU rows observed: 226
  - Current queued (server_state IN 1,2,4) for GPU prefixes: 27
- GPU host list observed for those prefixes:
  287(DESKTOP-N5RAJSE), 23(jisoo), 1(Pyhelix), 339(Foxtrot-2), 341(DESKTOP-ELBSBOI), 327(Echo-3), 16(dahyun), 29(DESKTOP-P57624Q), 330(Delta-1), 159(achernar), 340(Foxtrot-3), 347(sirius), 9(dbgrensenh27), 324(Charlie-1), 319(Dell-XPS-15-9560), 332(Delta-2), 353(Thing0L_4000), 321(Rosie).

NEW EXPERIMENTS DESIGNED + NOVELTY CHECK DOCUMENTATION
- New CPU experiment script added: wd_batchnoise_interaction.py.
- Scientific hypothesis tested: whether late weight-decay benefit is stronger under small-batch gradient-noise conditions (interaction effect), relative to large-batch conditions.
- Novelty-check evidence captured in run log includes explicit external search queries before script creation, including:
  1) "weight decay batch size interaction neural networks"
  2) "arxiv weight decay batch size interaction deep learning"
  3) "Scheduled Weight Decay paper arxiv 2021"
  4) site:arxiv.org searches for label smoothing/weight decay and related interaction literature.
- Script uploaded to /opt/axiom_boinc/html/user/experiments/wd_batchnoise_interaction.py and py_compile check returned OK.

KEY SCIENTIFIC FINDINGS
1. No new validated result artifact in this session reverses the established inverse critical-period weight-decay timing conclusion.
2. Validation throughput remains constrained by missing upload artifacts (0/240 sampled uncredited completed rows had payloads), so operational credit fairness continues to dominate this cycle.
3. A new mechanistic experiment line (wd_batchnoise_interaction) was introduced to test a specific interaction-level hypothesis: whether small-batch gradient noise amplifies late-WD gain relative to large-batch settings.
4. GPU mechanistic lines (wd_curvature_trigger_gpu / wd_timing_scale_gpu) remain active across a multi-host GPU cohort, preserving capacity to resolve curvature-gated vs fixed-onset timing behavior in upcoming result-bearing sessions.

NOTES
- No cumulative result-ID lists are included here by design; database state remains source of truth for credited rows.